"AI is the new electricity," Andrew Ng, chief scientist at Baidu Research, told Fortune. Ng predicts that the advent of AI-related technologies will transform all industries, as electricity did more than 100 years ago.
Indeed, AI and machine learning seem to be the hottest technologies in healthcare. Heather Mack at MobiHealthNews shares data supporting this:
- Many AI-focused healthcare companies raised their first equity rounds after January 2015.
- The number of AI healthcare deals grew more than three-fold from 2012 to 2016.
- More than a third of hospitals plan to adopt AI within two years.
Despite this fast growth, there are still a multitude of challenges facing the adoption of machine learning specifically, and AI in general, in healthcare. In this article, we examine some of these issues.
Managing and Integrating Large Data Sets
Machine learning has no utility unless it is fed datasets from which to learn. In the healthcare field, the availability of data is not an issue. Thanks to the HITECH component of the American Recovery and Reinvestment Act (ARRA), healthcare data is now readily available both in structured and unstructured formats.
The issue, Erin Dietsche at MedCityNews says, is that the data is in silos. Author Sebastian Raschka agrees and suggests that the problem can be attributed to how "[healthcare] data is highly heterogenous, and cleaning and combining data from different databases is probably the bottle neck." Dietsche also cites budget constraints on the part of providers, as it is costly to not only maintain current technologies but also invest in new ones.
Jennifer Bresnick at Health IT Analytics suggests that the use of semantic data lakes can offer the flexibility needed to collect and store large amounts of data that can be used to power machine learning. This, Bresnick says, is due to the "unique ability [of data lakes] to synthesize and normalize disparate datasets and draw conclusions from seemingly unrelated pieces of information." The use of data lakes still requires a human touch. Curators are needed to ensure the accuracy, uniformity, and completeness of the data being added to the system.
Related to the issue of integrating large data sets is interoperability. To reap the full benefits of machine learning, it's crucial to synthesize medical and patient data into one system that is accessible to all providers in real-time.
One solution is to migrate all healthcare data into one system that can meet the needs of providers, payers, and patients. However, given the complexity of the healthcare system and its many moving parts, this can be a tall order. A more practical approach would be creating interoperability between the existing systems.
Interoperability will allow the various systems to "speak" with each other and enable providers to keep using their current systems. Providers can use a mix of different systems to manage different functions, as suit their needs, while allowing medical and patient data to be shared with others within the healthcare system. This effectively creates a single database that machine learning can work on in order to provide predictive analytics that can help patient care and improve health outcomes.
However, interoperability has been a difficult goal to achieve. Bresnick, in another piece, calls it a "perennial concern" and cites "fundamental differences in the way electronic health records are designed and implemented" as a major stumbling block. Unless interoperability can be attained, it's arguable that the full potential of using machine learning in healthcare will be stunted, as well.
Protecting Data Security and Patient Privacy
The collection of large datasets, particularly those concerning an individual's medical history, necessarily raises the question of privacy. Currently, there are legal limitations on access to medical data. This restricted access is a barrier to developing robust health-focused algorithms.
However, individual privacy is a genuine concern when it comes to data collection. Stephen Gardner at Bloomberg BNA reports on how international privacy regulators are looking into the ways machine learning and other technologies can affect privacy.
Areas of concern include:
- the ambiguity of data collection practices,
- lack of knowledge about how machine learning will use or reuse the data,
- and the question of accountability for automated decision making.
There are no ready solutions, although there is healthy debate about how to resolve these issues as machine learning technology becomes more proliferated.
Similarly, data security is also a concern. Breach Report 2016: Protected Health Information (PHI) reveals that 81 percent of breached records in 2016 came from hacking attacks, and that there were 335 large-scale personal health information (PHI) data breaches compromising 16,612,985 individual patient records. The security of electronic health records and medical data needs to be improved so that all parties feel more confident in storing and sharing data digitally.
Michael Bruemmer at MedCityNews has three recommendations for improving data security. He suggests scaling up training for staff who handle EHRs to minimize any data breaches by employees. Aside from ensuring basic security protocols are in place, it is important to educate staff on data security policies and procedures and train them to spot signs of security threats. Healthcare organizations should also have a data breach response plan in place and practice the implementation of the plan with their staff. Lastly, investment in security is a must.
Getting 'Truth' From Machine Learning
In an interview with The Mission, Dr. Dave Channin argues that the real issue with data as it pertains to machine learning is "truth." He refers to medical imaging in particular, and the importance of "knowing what is in the image."
The complexity of medical images makes the annotation or validation of data very challenging. Consider these factors: numerous medical imaging devices in the market, thousands of observable imaging features, only 35,000 professionals in the United States trained to annotate, and the need for consensus annotations by multiple expert observers to minimize human error. This lack of granularity is a challenge even for unsupervised machine learning.
Finding The Balance Between Technology and the Human Touch
As Puneet Gupta at Harvard Science Review puts it, "Machine learning brings about a heated debate on ethics." At the heart of this ethical debate is one big question: To what extent will machine learning replace doctors?
Broadly speaking, most in the field see machine learning as a clinical support tool, as seen in Laura Dyrda's piece at Becker's Hospital Review. Mudit Garg, co-founder and CEO of Qventus, tells Dyrda that advances in machine learning will have the effect of lightening the load on healthcare professionals. "They will be able to focus solely on those issues that require their attention and spend the rest of their time dedicating themselves to their patients," he says.
Lisa Suennen, managing director at GE Ventures, agrees and states that machine learning "allows clinicians to work at the highest level of their ability by making them far more informed and effective patient advocates."
The complex and ever-evolving nature of healthcare means that human doctors will remain a key part of our healthcare system, regardless of technological advances. "At some point, human judgment is a lot more valuable than any insights AI can provide," Kapila Ratnam at MedCityNews argues.
As machine learning capabilities are further developed, there could come a point where technology can detect the onset of diseases before they actually manifest. At a point in diagnosis when human tissue and cells may be classed as "indeterminate," doctors can play an important role in deciding whether or not to deliver care.
Ratnam brings up the salient point that studies have shown that often it is best to leave the human body to heal itself without medical intervention, particularly in the early phase of any disease. The bottom line is that there is little room for error when it comes to healthcare, and healthcare professionals are the last line of defense against "artificial stupidity."
Mass vs. Individual
Regardless of how machine learning capabilities in healthcare develop in the future, Bill Simpson at MedCityNews cautions against pinning all our hopes on machine learning. Simpson is of the view that machine learning is not the end all for healthcare because "the [healthcare] system is highly complex and humans are inherently irrational beings."
In his opinion, a personal rather than population view of patient engagement would be more effective. He posits this idea: "What if servers and algorithms were not interested in 1 model to predict the action of 100,000 people, but 100,000 models, to predict the actions of each individual person. What if your data was analyzed over time so that machine learning could give you insights into your own self?"
Giving individuals predictive insight on their own behavior, Simpson believes, would be a more empowering approach. It's certainly something to ponder.
Images by: kentoh/©123RF Stock Photo, jirsak/©123RF Stock Photo, kentoh/©123RF Stock Photo, megaflopp/©123RF Stock Photo
Tags: Artificial Intelligence